AI's Tower of Babel: Why 'Harness' and 'Scaffold' Mean Everything and Nothing

If you’ve ever nodded along while secretly wondering what separates an LLM ‘harness’ from its ‘scaffold,’ you’re in crowded company. After ICLR 2026, researcher Aritra Roy Gosthipaty captured the collective confusion perfectly, asking why the same terms sparked a dozen different explanations. A new glossary from Chintala and Gosthipaty doesn’t try to crown a single winner, but it does draw a sharp, practical line that cuts through the noise.

The model is just the brain—Claude, Qwen, GPT—a text-in, text-out engine with no memory. To become an agent that actually does things, it needs a harness. This is the execution loop that calls the model, manages tool calls, handles errors, and decides when the job is done. Think of it as the agent’s engine. Products like Claude Code are often called a harness outright; its own docs state, ‘Claude Code serves as the agentic harness around Claude.’

The scaffolding, by contrast, is the behavior-defining layer the model works from: the system prompt, tool descriptions, and the format for parsing its responses. It’s the worldview, while the harness is the machinery that acts on it. The distinction matters most during training pipelines, where you tune the scaffold’s instructions separately from the harness’s execution logic. Confusingly, you’ll also hear ‘scaffold’ used broadly for any infrastructure the harness relies on.

The glossary also clarifies where an orchestrator fits in. It’s a higher-level boss that coordinates multiple agents as units, each running its own harness—a useful distinction when scaling beyond a single loop. For practitioners, the core takeaway from the community is blunt: if you’re not the model, you’re the harness. Everything else is a detail you can finally stop arguing about.